talesuite/tale_suite_trajectories_downsampled

Text-Adventure Agent Trajectories

Anonymous for NeurIPS 2026 Evaluations and Benchmark Track.

Agent trajectory data from a text-adventure benchmark suite.

Note: Only top 10 model trajectories included for this downsampled repo

Leaderboard

Top agents ranked by average best normalized score per game across 122 games, each repeated over 5 seeds (610 total). Scores reflect the highest normalized score achieved at any point during each playthrough, not the final score.

Rank	Model	TW	TWX	ALFWorld	SciWorld	Jericho	benchmark
1	claude-opus-4.6 (high)	1.000	0.957	1.000	0.920	0.162	0.604
2	claude-opus-4.5 (high)	1.000	0.891	1.000	0.940	0.168	0.603
3	o3 (medium)	1.000	0.919	0.883	0.930	0.157	0.587
4	gpt-5.1 (high)	0.995	0.898	0.917	0.902	0.161	0.582
5	o3 (high)	1.000	0.896	0.817	0.931	0.161	0.580
6	claude-sonnet-4.6 (high)	1.000	0.905	1.000	0.887	0.130	0.575
7	gpt-5 (high)	1.000	0.755	0.933	0.918	0.172	0.575
8	o3 (low)	0.991	0.898	0.700	0.883	0.142	0.548
9	claude-4-sonnet	0.996	0.784	0.917	0.870	0.123	0.543
10	claude-3.7-sonnet (1024)	0.973	0.913	0.833	0.765	0.125	0.525

Overview

This dataset contains agent game trajectories collected across multiple text-adventure game frameworks. Each trajectory records a full episode of an LLM-driven agent interacting with a text-based game environment, including observations, actions, scores, and (where available) thinking traces.

Frameworks

• alfworld
• jericho
• scienceworld
• textworld
• textworldexpress

Dataset Structure

Each row is a single trajectory with the following fields:

Field	Description
transcript_id	Unique identifier for the trajectory
framework	Game framework (e.g., textworld, jericho)
game	Specific game name
model	LLM used to drive the agent
agent_type	Agent architecture (e.g., zero-shot)
score	Final score achieved
max_score	Maximum possible score
normalized_score	Score normalized to [0, 1]
total_steps	Number of agent actions taken
turns	Full conversation history (JSON string)
score_progression	Score at each step (JSON string)

Usage

from datasets import load_dataset

# Load a specific framework
ds = load_dataset("talesuite/tale_suite_trajectories_downsampled", "alfworld")

# Load all frameworks
for fw in ["alfworld", "jericho", "scienceworld", "textworld", "textworldexpress"]:
    ds = load_dataset("talesuite/tale_suite_trajectories_downsampled", fw)
    print(f"{fw}: {len(ds['train'])} trajectories")

Raw Data

Raw JSONL files (one line per trajectory, metadata flattened) are available under raw/ for direct download.

Citation

Citation information will be added after the review process.

License

Please refer to the individual game frameworks for their respective licenses.